Comparative analysis of texture classification based on low and high order local features

The importance of texture for recognition of objects, scenes and events is well-known and used in various computer vision tasks. Until recently, best-performing texture classification algorithms relied on processing of low-level local features and statistical learning based adjustment of classifiers. Convolutional neural networks introduced higher order local features and improved classification results significantly. In this paper, we compared texture classification based on low-lever and high order local features. Also, we demonstrated the ability of convolutional networks to learn high order features from one dataset and to efficiently use that knowledge on a different dataset.